CN113254549A - Character relation mining model training method, character relation mining method and device - Google Patents
Character relation mining model training method, character relation mining method and device Download PDFInfo
- Publication number
- CN113254549A CN113254549A CN202110682820.3A CN202110682820A CN113254549A CN 113254549 A CN113254549 A CN 113254549A CN 202110682820 A CN202110682820 A CN 202110682820A CN 113254549 A CN113254549 A CN 113254549A
- Authority
- CN
- China
- Prior art keywords
- embedding
- entity
- initial
- quaternion
- time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/28—Databases characterised by their database models, e.g. relational or object models
- G06F16/284—Relational databases
- G06F16/288—Entity relationship models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
- G06N5/022—Knowledge engineering; Knowledge acquisition
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Animal Behavior & Ethology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a character relation mining model training method, a character relation mining method and a character relation mining device, wherein the training method comprises the following steps: acquiring a space-time knowledge map; randomly sampling a positive sample of the spatio-temporal knowledge map to generate a negative sample, and determining initial embedding of head entities, initial embedding of relations, initial embedding of tail entities and time embedding of the positive sample and the negative sample; carrying out vector rotation on the initial embedding of the head entity and the initial embedding of the tail entity to obtain the quaternion embedding of the head entity and the quaternion embedding of the tail entity; replacing the head entity initial embedding and the tail entity initial embedding with corresponding head entity quaternion embedding and tail entity quaternion embedding respectively to obtain a processed positive sample and a processed negative sample; and iteratively training the character relation mining model to be convergent by adopting the processed positive sample and the processed negative sample. The technical scheme of the invention can mine the relationship between the entities evolving along with the time change and improve the integrity of the knowledge graph.
Description
Technical Field
The invention relates to the technical field of knowledge graphs, in particular to a character relation mining model training method, a character relation mining method and a character relation mining device.
Background
The knowledge map is a series of different graphs for displaying the relation between the knowledge development process and the structure, describes knowledge resources and carriers thereof by using a visualization technology, and mines, analyzes, constructs, draws and displays knowledge and the mutual relation between the knowledge resources and the carriers. The space-time knowledge graph is added into the entity to form space attribute information, and management and query of space-time data in a specific target field can be achieved. The spatio-temporal knowledge map comprises a large number of facts containing spatio-temporal information, and each fact respectively forms a quadruple (s, r, o, t) comprising a head entity s, a relation r, a tail entity o and a time t. The space-time knowledge graph usually has the problem of link loss, which limits the application of the space-time knowledge graph, and in order to exert the functions of the space-time knowledge graph as much as possible, the space-time knowledge graph is usually required to be complemented, and is further expanded, so that the space-time knowledge graph is more complete, and the mutual relation between entities is disclosed.
At present, a knowledge graph embedding method is often adopted for performing knowledge graph completion, unknown entities and relationships between the entities are predicted according to the known entities and relationships between the entities, the methods can focus on the static relationships between the entities, but the relationships between a plurality of entities are not invariable, for example, the relationships between people are changed along with the change of time, and the existing knowledge graph completing method cannot predict the change of the relationships between the entities along with the change of time.
Disclosure of Invention
The invention solves the problem of how to mine the relationship between entities evolving along with the time change and improve the integrity of the knowledge graph.
In order to solve the above problems, the present invention provides a training method for a character relationship mining model, a character relationship mining method, a character relationship mining device, and a storage medium.
In a first aspect, the present invention provides a training method for a character relationship mining model, including:
acquiring a pre-established spatio-temporal knowledge map related to character relations, wherein the spatio-temporal knowledge map comprises a plurality of positive samples;
randomly sampling according to the positive samples to generate negative samples, and respectively determining head entity initial embedding, relation initial embedding, tail entity initial embedding and time embedding of each positive sample and each negative sample;
vector rotation is respectively performed on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding, and the vector rotation method comprises the following steps: the initial embedding of the head entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the head entity, the initial embedding of the tail entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the tail entity;
replacing the head entity initial embedding and the tail entity initial embedding in the positive sample and the negative sample with the corresponding head entity quaternion embedding and tail entity quaternion embedding respectively to obtain a processed positive sample and a processed negative sample;
and iteratively training a pre-established character relation mining model to be convergent by adopting the processed positive sample and the processed negative sample.
Optionally, the iteratively training the pre-established character relationship mining model to converge by using the processed positive sample and the processed negative sample includes:
performing conjugate operation on the tail entity quaternion embedding to obtain the conjugate embedding of the tail entity quaternion embedding;
embedding the head entity quaternion, the conjugate embedding of the tail entity quaternion and the relationship into a score function of an input character relationship mining model, and respectively scoring the positive sample and the negative sample by adopting the score function to obtain a score value of the positive sample and a score value of the negative sample;
and inputting the score values of the positive samples and the score values of the negative samples into a loss function of the character relation mining model, and optimizing the character relation mining model by minimizing the loss function.
Optionally, the scoring function comprises:
wherein，
Is a sample
The score value of (a) is calculated,
is either a positive or a negative sample,
to represent
The head entity quaternion embedding of a time of day,
indicating the initial embedding of the relationship(s),
to represent
Conjugate embedding of the tail entity quaternion embedding of a time instant.
Optionally, before the obtaining of the pre-established spatio-temporal knowledge graph about human relationships, the method includes:
acquiring initial data about the relationship of people;
carrying out named entity identification on the initial data to obtain entities, and carrying out relation extraction on the initial data to obtain the relation between the entities;
and constructing the spatio-temporal knowledge graph according to the entity and the relation between the entities.
Optionally, the randomly sampling according to the positive samples comprises:
for any one of the positive samples, randomly selecting an entity from an entity set to replace a head entity or a tail entity of the positive sample, and obtaining one negative sample, wherein the entity set comprises all the head entities and the tail entities.
In a second aspect, the present invention provides a method for mining a character relationship, including:
acquiring three initial elements, wherein the initial elements comprise any three of a first person, a second person, a relationship and time;
taking each element in the set corresponding to the missing element as an undetermined element to form a quadruple with the three initial elements, wherein the quadruple comprises the first person, the second person, the relationship and the time;
preprocessing each quadruple to obtain a processed quadruple;
inputting each processed quadruple into a trained character relationship mining model, scoring each processed quadruple by adopting a scoring function, and determining the score value of each element to be determined, wherein the character relationship mining model is obtained by adopting the training method of the character relationship mining model;
and determining mining information according to the score value of each element to be determined.
Optionally, the preprocessing each quadruple comprises:
respectively determining initial embedding of a head entity, initial embedding of a relationship, initial embedding of a tail entity and time embedding of each quadruple;
respectively carrying out vector rotation on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding to obtain the quaternion embedding of the head entity and the quaternion embedding of the tail entity in a three-dimensional space;
and replacing the head entity initial embedding and the tail entity initial embedding in the quadruple respectively with the corresponding head entity quaternion embedding and the corresponding tail entity quaternion embedding to obtain the processed quadruple.
In a third aspect, the present invention provides a training apparatus for a character relationship mining model, including:
the acquisition module is used for acquiring a pre-established spatiotemporal knowledge map related to character relations, wherein the spatiotemporal knowledge map comprises a plurality of positive samples;
the sampling module is used for carrying out random sampling according to the positive samples to generate negative samples and respectively determining head entity initial embedding, relation initial embedding, tail entity initial embedding and time embedding of each positive sample and each negative sample;
a rotation module, configured to perform vector rotation on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding, respectively, including: the initial embedding of the head entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the head entity, the initial embedding of the tail entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the tail entity;
a replacing module, configured to replace the head entity initial embedding and the tail entity initial embedding in the positive sample and the negative sample with the corresponding head entity quaternion embedding and the corresponding tail entity quaternion embedding, respectively, so as to obtain a processed positive sample and a processed negative sample;
and the training module is used for iteratively training a pre-established character relation mining model to be convergent by adopting the processed positive sample and the processed negative sample.
In a fourth aspect, the present invention provides an information mining apparatus, including:
the system comprises an acquisition module, a display module and a display module, wherein the acquisition module is used for acquiring three initial elements, and the initial elements comprise any three of a first person, a second person, a relationship and time;
a configuration module, configured to use each element in a set corresponding to a missing element as an undetermined element to form a quadruple with the three initial elements, where the quadruple includes the first person, the second person, the relationship, and the time;
the processing module is used for preprocessing each quadruple to obtain a processed quadruple;
the scoring module is used for respectively inputting each processed quadruple into a trained character relationship mining model, scoring each processed quadruple by adopting a scoring function, and determining the score value of each element to be determined, wherein the character relationship mining model is obtained by adopting the training method of the character relationship mining model;
and the output module is used for determining the mined information according to the score values of the undetermined elements.
In a fifth aspect, the present invention provides an electronic device comprising a memory and a processor;
the memory for storing a computer program;
the processor is configured to implement the above-mentioned training method of the character relationship mining model or the above-mentioned character relationship mining method when executing the computer program.
In a sixth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, implements the training method of the character relationship mining model as described above or the character relationship mining method as described above.
The character relationship mining model training method, the character relationship mining device and the storage medium have the advantages that: the knowledge graph can be established in advance according to the acquired initial data about the character relationship, and negative samples are randomly generated according to positive samples in the space-time knowledge graph. Vector rotation is carried out on head entity initial embedding and tail entity initial embedding of the positive sample and the negative sample according to time embedding, time information is fused, embedding expressions of the head entity initial embedding and the tail entity initial embedding in a three-dimensional vector space are determined, head entity quaternion embedding and tail entity quaternion embedding of the fused space-time information are obtained, and therefore evolution of the relation between the entities in the space-time map and the entity changing along with time can be well predicted. The initial embedding of the head entity and the initial embedding of the tail entity in the positive sample and the negative sample are respectively replaced by corresponding embedding of the quaternion of the head entity and embedding of the quaternion of the tail entity, and the character relationship mining model is iteratively trained by adopting the positive sample and the negative sample after replacement, so that the method can be used for predicting the relationship between the entity and the entity which evolves along with time change, can complete the knowledge graph, and improves the integrity of the knowledge graph.
Drawings
FIG. 1 is a schematic flow chart of a training method of a character relationship mining model according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating the result of vector rotation for initial embedding of a header entity according to an embodiment of the present invention;
FIG. 3 is a schematic flow chart illustrating a method for mining relationships between people according to another embodiment of the present invention;
FIG. 4 is a schematic structural diagram of a training apparatus for a character relationship mining model according to yet another embodiment of the present invention;
fig. 5 is a schematic structural diagram of a character relationship mining device according to another embodiment of the present invention.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in detail below.
It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein.
The current method for embedding and complementing the spatiotemporal knowledge graph can only reflect the static relationship between entities, and cannot capture the change of the internal relationship between the entities and the relationship between the entities along with time, but in reality, the relationship between a plurality of entities and the relationship between the entities are not invariable and can change along with time, for example, famous actors of Kit Harington and Rose Leslie love in 2012, the relationship between two people is lovers, the two people marry in 2017, the relationship between the two people is quasi-couples, and then the two people marry in 2018, and the relationship between the two people becomes couples. The relationship between the couple and the couple is a common relationship evolution process in real life, and is suitable for a plurality of couples.
As shown in fig. 1, in the training method for a character relationship mining model provided in the embodiment of the present invention, the character relationship mining model is a knowledge graph embedded model, and the training method for the character relationship mining model is a training method for the knowledge graph embedded model, and the training method includes:
step S110, a pre-established spatio-temporal knowledge map about the relationship of the characters is obtained, and the spatio-temporal knowledge map comprises a plurality of positive samples.
Specifically, the spatio-temporal knowledge graph comprises a plurality of positive samples, each positive sample is represented by a quadruple, each quadruple comprises a head entity, a relation, a tail entity and time, and for the spatio-temporal knowledge graph related to the relationship of the characters, the positive samples comprise a first character, the relationship between the characters, a second character and the time, wherein the entities are the characters, and the relationship is the relationship between the characters.
Optionally, before the obtaining of the pre-established spatio-temporal knowledge graph about human relationships, the method includes:
acquiring initial data about the relationship of people;
carrying out named entity identification on the initial data to obtain entities, and carrying out relation extraction on the initial data to obtain the relation between the entities;
and constructing the spatio-temporal knowledge graph according to the entity and the relation between the entities.
Specifically, the required relevant initial data can be collected in the internet through a data crawling tool, for example, when a spatiotemporal knowledge map about the relationship of people is established, the initial data of relevant people is crawled in the internet, and the initial data comprises the information of people and the relationship between people. And then, adopting a Stanford-NLP (natural language processing tool) to perform named entity identification and relationship extraction to obtain the relationship between the entities, and constructing a corresponding spatiotemporal knowledge graph according to the relationship between the entities, namely identifying and extracting the relationship between the characters in the initial data and constructing the spatiotemporal knowledge graph related to the relationship between the characters.
In the optional embodiment, initial data is obtained, and a spatiotemporal knowledge graph is established according to the initial data, so that the relation between the entities can be intuitively reflected, and information is conveniently retrieved and mined.
And step S120, carrying out random sampling according to the positive samples to generate negative samples, and respectively determining the initial embedding of the head entity, the initial embedding of the relationship, the initial embedding of the tail entity and the time embedding of each positive sample and each negative sample.
Specifically, initial embedding of a head entity corresponding to the head entity is represented in a quaternion form, initial embedding of a relationship corresponding to the relationship and initial embedding of a tail entity corresponding to the tail entity are represented, and time embedding corresponding to time is represented by a unit quaternion. The quaternion can be expressed as q = a + bi + cj + dk, where q is a quaternion, where a, b, c, and d are real numbers, i, j, and k are imaginary units, and are also imaginary space coordinate axes of the quaternion, and a specific expression form is the prior art, which is not described herein again.
Optionally, the randomly sampling according to the positive samples comprises:
for any one of the positive samples, randomly selecting an entity from an entity set to replace a head entity or a tail entity of the positive sample, and obtaining one negative sample, wherein the entity set comprises all the head entities and the tail entities.
Specifically, for a positive sample (head entity, relationship, tail entity, time), when the positive sample in the spatio-temporal knowledge graph about the relationship between the people is (first person, relationship between the people, second person, time), a corresponding negative sample is generated, the head entity or the tail entity is covered, and an entity is randomly selected from the entity set consisting of all the entities, the covered head entity or tail entity in the positive sample is replaced, a negative sample is generated, and the negative sample is (randomly selected entity, relationship, tail entity, time) or (head entity, relationship, randomly selected entity, time), accordingly, the negative sample corresponding to the positive sample in the spatio-temporal knowledge graph about the relationship between the people can be (randomly selected person, relationship between the people, second person, time) or (first person, relationship between the people, randomly selected person, time).
In the optional embodiment, the corresponding negative sample is generated by replacing the head entity or the tail entity in the positive sample, so that the generation speed of the negative sample can be increased, and the method is simple and efficient.
And step S130, respectively carrying out vector rotation on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding, and obtaining the quaternion embedding of the head entity and the quaternion embedding of the tail entity in a three-dimensional space.
The vector rotation of the initial embedding of the head entity and the initial embedding of the tail entity respectively according to the time embedding comprises:
the initial embedding of the head entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the head entity;
and carrying out initial embedding on the tail entity, and carrying out left multiplication on the time embedding, and carrying out right multiplication on the inverse of the time embedding to obtain the quaternion embedding of the tail entity.
Specifically, taking initial embedding of a head entity as an example, vector rotation of the initial embedding of the head entity and the initial embedding of the tail entity is explained.
As shown in FIG. 2, assume that the initial embedding of the head entity is V and the time embedding expressed by unit quaternion is V
，
Wherein
Is a unit vector in a three-dimensional vector space, the head entity is initially embedded
About an axis of rotation
Rotation angle
Is represented by a first formula comprising:
wherein the content of the first and second substances,
the vector obtained after the initial embedding rotation of the head entity, namely the quaternion embedding of the head entity,
for time embedding
Conjugation of (1).
Since, according to the norm formula of the quaternion,
，
and, the inverse of quaternion
By
Definition, multiplication by
Then obtain
From which follows:
for a unit quaternion, the number of bits in the unit,
then, then
Expanding the first formula to obtain
Rule of product of two imaginary vectors
And the method can further obtain the result that,
using trigonometric identities, one can obtain
Wherein the content of the first and second substances,
and
is the initial embedding of the head entity
Are perpendicular and parallel to the axis, respectively
Which satisfies the Rodrigues rotation formula in the three-dimensional vector space, the rotation of the head entity initial embedding and the tail entity initial embedding in the three-dimensional vector space can be determined by the first formula.
In the optional embodiment, the initial embedding of the head entity and the initial embedding of the tail entity are subjected to vector rotation, the vector rotation is carried out based on the time embedding, the time information is fused, the representation of the initial embedding of the head entity and the initial embedding of the tail entity in the three-dimensional vector space is determined, the quaternion embedding of the head entity and the quaternion embedding of the tail entity which are fused with the time-space information are obtained, the evolution of the relation between the entities and the entities along with the time transformation can be well predicted, the relation between the entities after the evolution is conveniently mined, the spatiotemporal knowledge map is supplemented, and the integrity of the spatiotemporal knowledge map is improved.
Step S140, replacing the head entity initial embedding and the tail entity initial embedding in the positive sample and the negative sample with the corresponding head entity quaternion embedding and the corresponding tail entity quaternion embedding, respectively, to obtain a processed positive sample and a processed negative sample.
Specifically, it is assumed that the positive samples and the negative samples include head entity initial embedding, relationship initial embedding, tail entity initial embedding and time embedding, and after the replacement, the head entity quaternion embedding, relationship initial embedding, tail entity quaternion embedding and time embedding are performed.
And S150, iteratively training a pre-established character relation mining model by adopting the processed positive sample and the processed negative sample until convergence.
Optionally, the character relationship mining model is a knowledge graph embedding model, and includes a score function and a loss function, and iteratively training the pre-established character relationship mining model to converge by using the processed positive sample and the processed negative sample includes:
and step S151, carrying out conjugate operation on the tail entity quaternion embedding to obtain the conjugate embedding of the tail entity quaternion embedding.
Specifically, the conjugate of the tail entity quaternion embedding is solved, the conjugate of the quaternion and the quaternion have the same real part and an imaginary part with opposite sign, and the conjugate of the tail entity quaternion embedding is assumed to be a + bi + cj + dk, wherein a, b, c and d are real numbers, and i, j and k are imaginary number units, so that the conjugate of the tail entity quaternion embedding is a-bi-cj-dk.
Step S152, embedding the head entity quaternion, embedding the conjugate of the tail entity quaternion and initially embedding the relationship into a score function of an input character relationship mining model, and respectively scoring the positive sample and the negative sample by adopting the score function to obtain the score value of the positive sample and the score value of the negative sample.
Optionally, the scoring function comprises:
wherein the content of the first and second substances,
is a sample
The score value of (a) is calculated,
is either a positive or a negative sample,
to represent
The head entity quaternion embedding of a time of day,
indicating the initial embedding of the relationship(s),
to represent
Conjugate embedding of the tail entity quaternion embedding of a time instant.
Specifically, embedding the quaternion of the head entity and the initial embedding of the relationship are added, then the conjugate embedding of the quaternion of the tail entity is subtracted to obtain a calculation result, and modular length operation is carried out on the calculation result to obtain a score value corresponding to the sample.
Step S153, inputting the score values of the positive samples and the score values of the negative samples into a loss function of the character relationship mining model, and optimizing the character relationship mining model by minimizing the loss function.
Specifically, a MarginLoss function may be adopted based on the negative sample to adjust parameters of the character relationship mining model, so that the score value of the positive sample obtained through calculation is as small as possible, and the score value of the negative sample is as large as possible, and the MarginLoss function is the prior art and is not described herein again. In order to improve the training speed, the iterative training can be stopped when the value of the loss function is smaller than a preset threshold value, and a trained character relation mining embedded model is obtained.
In this embodiment, the knowledge graph may be pre-established according to the initial data about the relationship of the persons, and the negative samples may be randomly generated according to the positive samples in the spatio-temporal knowledge graph. Vector rotation is carried out on head entity initial embedding and tail entity initial embedding of positive samples and negative samples according to time embedding, time information is fused, embedding representation of the head entity initial embedding and the tail entity initial embedding in a three-dimensional vector space is determined, head entity quaternion embedding and tail entity quaternion embedding of fused space-time information can be obtained, and therefore evolution of the relation between entities in a space-time map and the relation between the entities along with time change can be well predicted. The initial embedding of the head entity and the initial embedding of the tail entity in the positive sample and the negative sample are respectively replaced by corresponding embedding of the quaternion of the head entity and embedding of the quaternion of the tail entity, and the character relationship mining model is iteratively trained by adopting the positive sample and the negative sample after replacement, so that the method can be used for predicting the relationship between the entity and the entity which evolves along with the change of time, can complement the entity and the relationship which are missing and wrong in the knowledge graph, and improves the integrity of the knowledge graph.
The method fully utilizes the advantages of quaternion in expression capability, carries out embedded expression on each element of the knowledge map, and gives play to the representation advantages of quaternion theory in three-dimensional vector space rotation to model the change of the entities and the relation along with time, thereby being convenient for fully mining the dynamic evolution of the mutual relation between the entities in real time. For example, for the relationship of the characters, the evolution of the relationship between the characters along with the time can be accurately determined, the hidden mutual relationship and suspicious relationship between the characters are mined, the completion of the space-time knowledge map is efficiently realized, the existing character relationship space-time knowledge map is enriched and expanded, the integrity and the application range of the space-time knowledge map are improved, and the method can be used for discovering and preventing fraud crimes.
Compared with the prior art, the method has better effect in the experimental task of predicting the entity or the relationship between the entities. For example, in 4 public data sets ICEWS14, ICEWS05-15, YAGO11k and GDELT, the invention achieves the best effect on four knowledge map evaluation indexes of MRR, Hits @10, Hits @3 and Hits @1 through practice tests, and the data index distribution variance is small and the performance is stable.
As shown in fig. 3, a person relationship mining method according to another embodiment of the present invention is an information mining method, and includes:
step S210, three initial elements are obtained, wherein the initial elements comprise any three of a first person, a second person, a relationship and time, the first person is a first entity, the second person is a second entity, and the relationship is an inter-person relationship;
step S220, each element in the set corresponding to the missing element is respectively used as an undetermined element to form a quadruple with the three initial elements, where the quadruple includes the first person, the second person, the relationship, and the time.
Specifically, the entity set includes all the first entities and the second entities, that is, includes all the people, the relationship set includes all the relationships, the time set includes all the times, and when the three initial elements are the first entities, the second entities and the relationships, each element in the time set is respectively used as an undetermined element to form a quadruple with the three initial elements; when the three initial elements are a first entity, a second entity and time, taking each element in the relation set as an undetermined element to form a quadruple with the three initial elements; and when the three initial elements are the first entity, the relation and the time, or the second entity, the relation and the time, taking each element in the entity as an undetermined element to form a quadruple with the three initial elements.
And step S230, preprocessing each quadruple to obtain a processed quadruple.
Optionally, the preprocessing each quadruple comprises:
respectively determining initial embedding of a head entity, initial embedding of a relationship, initial embedding of a tail entity and time embedding of each quadruple;
respectively carrying out vector rotation on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding to obtain the quaternion embedding of the head entity and the quaternion embedding of the tail entity in a three-dimensional space;
and replacing the head entity initial embedding and the tail entity initial embedding in the quadruple respectively with the corresponding head entity quaternion embedding and the corresponding tail entity quaternion embedding to obtain the processed quadruple.
Step S240, inputting each processed quadruple into a trained character relationship mining model, scoring each processed quadruple by adopting a scoring function, and determining the score value of each element to be determined, wherein the character relationship mining model is obtained by adopting the training method of the character relationship mining model;
and step S250, determining mining information according to the score value of each undetermined element.
Specifically, the undetermined element with the lowest score is output as the mined information, if the three initial elements obtained are (a, a couple,/, 2021), that is, the initial elements include an entity, a relationship and a time, each entity in the entity set is respectively taken as the undetermined element to form a quadruple with the initial elements, if the entity in the entity set includes b, c and d in addition to the a, each entity except the a is respectively taken as three quadruples with the three initial elements, that is, (a, a couple, b, 2021), (a, a couple, c, 2021) and (a, a couple, d, 2021), the three quadruples are respectively input into the trained character relationship mining model, after scoring through the scoring function, the score of b is assumed to be 0.01, the score of c is 0.4, the score of d is 0.5, the entity with the lowest score "b" is output as the predicted entity, i.e., mined information.
In the embodiment, three initial elements are obtained, and are supplemented through undetermined elements in an entity set, a relation set or a time set to form a plurality of quadruples, each quadruple is preprocessed and converted into an embedded representation, the embedded representation is input into a trained character relation mining model, score values of the quadruples are determined through a scoring function, and the undetermined element corresponding to the quadruple with the lowest score value is output as mining information, so that the method is simple and efficient, and can be used for mining the entity and the relation between the entities which evolve along with the time change, and the integrity of the space-time knowledge map is greatly improved.
As shown in fig. 4, a training apparatus for a character relationship mining model according to still another embodiment of the present invention is a training apparatus for a knowledge-graph-embedded model, and includes:
the acquisition module is used for acquiring a pre-established spatiotemporal knowledge map related to character relations, wherein the spatiotemporal knowledge map comprises a plurality of positive samples;
the sampling module is used for carrying out random sampling according to the positive samples to generate negative samples and respectively determining head entity initial embedding, relation initial embedding, tail entity initial embedding and time embedding of each positive sample and each negative sample;
a rotation module, configured to perform vector rotation on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding, respectively, including: the initial embedding of the head entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the head entity, the initial embedding of the tail entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the tail entity;
a replacing module, configured to replace the head entity initial embedding and the tail entity initial embedding in the positive sample and the negative sample with the corresponding head entity quaternion embedding and the corresponding tail entity quaternion embedding, respectively, so as to obtain a processed positive sample and a processed negative sample;
and the training module is used for iteratively training a pre-established character relation mining model to be convergent by adopting the processed positive sample and the processed negative sample.
As shown in fig. 5, a person relationship mining device according to still another embodiment of the present invention is an information mining device, including:
the system comprises an acquisition module, a display module and a display module, wherein the acquisition module is used for acquiring three initial elements, the initial elements comprise any three of a first person, a second person, a relationship and time, the first person is a first entity, and the second person is a second entity;
a configuration module, configured to use each element in a set corresponding to a missing element as an undetermined element to form a quadruple with the three initial elements, where the quadruple includes the first person, the second person, the relationship, and the time;
the processing module is used for preprocessing each quadruple to obtain a processed quadruple;
the scoring module is used for respectively inputting each processed quadruple into a trained character relationship mining model, scoring each processed quadruple by adopting a scoring function, and determining the score value of each element to be determined, wherein the character relationship mining model is obtained by adopting the training method of the character relationship mining model;
and the output module is used for determining the mined information according to the score values of the undetermined elements.
Another embodiment of the present invention provides an electronic device, including a memory and a processor; the memory for storing a computer program; the processor is configured to, when executing the computer program, implement the above-described method for training a character relationship mining model or the above-described method for character relationship mining, that is, implement a corresponding method for training a knowledge-graph embedded model or a corresponding method for information mining. The electronic device may be a computer or a server, etc.
Yet another embodiment of the present invention provides a computer readable storage medium having stored thereon a computer program which, when being executed by a processor, implements a method for training a personality relationship mining model as described above or a method for persona relationship mining as described above, i.e., implements a method for training a corresponding knowledge-graph-embedded model or a corresponding method for information mining.
It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by a computer program, which can be stored in a computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), or the like. In this application, the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment of the present invention. In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.
The processor may be an integrated circuit chip having signal processing capabilities. The Processor may be a general-purpose Processor, and includes a Central Processing Unit (CPU), a Network Processor (NP), and the like. The Processor may also be other general purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The various methods, steps, and logic blocks disclosed in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
Although the present disclosure has been described above, the scope of the present disclosure is not limited thereto. Various changes and modifications may be effected therein by one of ordinary skill in the pertinent art without departing from the spirit and scope of the present disclosure, and these changes and modifications are intended to be within the scope of the present disclosure.
Claims (10)
1. A training method for a character relation mining model is characterized by comprising the following steps:
acquiring a pre-established spatio-temporal knowledge map related to character relations, wherein the spatio-temporal knowledge map comprises a plurality of positive samples;
randomly sampling according to the positive samples to generate negative samples, and respectively determining head entity initial embedding, relation initial embedding, tail entity initial embedding and time embedding of each positive sample and each negative sample;
vector rotation is respectively performed on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding, and the vector rotation method comprises the following steps: the initial embedding of the head entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the head entity, the initial embedding of the tail entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the tail entity;
replacing the head entity initial embedding and the tail entity initial embedding in the positive sample and the negative sample with the corresponding head entity quaternion embedding and tail entity quaternion embedding respectively to obtain a processed positive sample and a processed negative sample;
and iteratively training a pre-established character relation mining model to be convergent by adopting the processed positive sample and the processed negative sample.
2. The method of claim 1, wherein iteratively training a pre-established character relationship mining model to converge using the processed positive samples and the processed negative samples comprises:
performing conjugate operation on the tail entity quaternion embedding to obtain the conjugate embedding of the tail entity quaternion embedding;
embedding the head entity quaternion, the conjugate embedding of the tail entity quaternion and the relationship into a score function of an input character relationship mining model, and respectively scoring the positive sample and the negative sample by adopting the score function to obtain a score value of the positive sample and a score value of the negative sample;
and inputting the score values of the positive samples and the score values of the negative samples into a loss function of the character relation mining model, and optimizing the character relation mining model by minimizing the loss function.
3. The method of claim 2, wherein the scoring function comprises:
wherein the content of the first and second substances,
is a sample
The score value of (a) is calculated,
is either a positive or a negative sample,
to represent
The head entity quaternion embedding of a time of day,
indicating the initial embedding of the relationship(s),
to represent
Conjugate embedding of the tail entity quaternion embedding of a time instant.
4. The method for training a human relationship mining model according to any one of claims 1 to 3, wherein before the obtaining of the pre-established spatio-temporal knowledge map about human relationships, the method comprises:
acquiring initial data about the relationship of people;
carrying out named entity identification on the initial data to obtain entities, and carrying out relation extraction on the initial data to obtain the relation between the entities;
and constructing the spatio-temporal knowledge graph according to the entity and the relation between the entities.
5. The method for training human relationship mining models according to any one of claims 1 to 3, wherein the randomly sampling according to the positive samples comprises:
for any one of the positive samples, randomly selecting an entity from an entity set to replace a head entity or a tail entity of the positive sample, and obtaining one negative sample, wherein the entity set comprises all the head entities and the tail entities.
6. A character relationship mining method is characterized by comprising the following steps:
acquiring three initial elements, wherein the initial elements comprise any three of a first person, a second person, a relationship and time;
taking each element in the set corresponding to the missing element as an undetermined element to form a quadruple with the three initial elements, wherein the quadruple comprises the first person, the second person, the relationship and the time;
preprocessing each quadruple to obtain a processed quadruple;
respectively inputting each processed quadruple into a trained character relationship mining model, scoring each processed quadruple by adopting a scoring function, and determining the score value of each element to be determined, wherein the character relationship mining model is obtained by adopting the training method of the character relationship mining model as claimed in any one of claims 1 to 5;
and determining mining information according to the score value of each element to be determined.
7. A training device for character relation mining models is characterized by comprising:
the acquisition module is used for acquiring a pre-established spatiotemporal knowledge map related to character relations, wherein the spatiotemporal knowledge map comprises a plurality of positive samples;
the sampling module is used for carrying out random sampling according to the positive samples to generate negative samples and respectively determining head entity initial embedding, relation initial embedding, tail entity initial embedding and time embedding of each positive sample and each negative sample;
a rotation module, configured to perform vector rotation on the initial embedding of the head entity and the initial embedding of the tail entity according to the time embedding, respectively, including: the initial embedding of the head entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the head entity, the initial embedding of the tail entity is multiplied by the time embedding at the left side, and is multiplied by the inverse of the time embedding at the right side to obtain the quaternion embedding of the tail entity;
a replacing module, configured to replace the head entity initial embedding and the tail entity initial embedding in the positive sample and the negative sample with the corresponding head entity quaternion embedding and the corresponding tail entity quaternion embedding, respectively, so as to obtain a processed positive sample and a processed negative sample;
and the training module is used for iteratively training a pre-established character relation mining model to be convergent by adopting the processed positive sample and the processed negative sample.
8. A character relationship mining apparatus, comprising:
the system comprises an acquisition module, a display module and a display module, wherein the acquisition module is used for acquiring three initial elements, and the initial elements comprise any three of a first person, a second person, a relationship and time;
a configuration module, configured to use each element in a set corresponding to a missing element as an undetermined element to form a quadruple with the three initial elements, where the quadruple includes the first person, the second person, the relationship, and the time;
the processing module is used for preprocessing each quadruple to obtain a processed quadruple;
a scoring module, configured to input each processed quadruple into a trained character relationship mining model, score each processed quadruple by using a scoring function, and determine a score value of each element to be determined, where the character relationship mining model is obtained by using the training method of the character relationship mining model according to any one of claims 1 to 5;
and the output module is used for determining the mined information according to the score values of the undetermined elements.
9. An electronic device comprising a memory and a processor;
the memory for storing a computer program;
the processor, when executing the computer program, is configured to implement the training method of the human relationship mining model according to any one of claims 1 to 5 or the human relationship mining method according to claim 6.
10. A computer-readable storage medium, characterized in that the storage medium has stored thereon a computer program which, when being executed by a processor, implements the method for training a character relationship mining model according to any one of claims 1 to 5 or the method for character relationship mining according to claim 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110682820.3A CN113254549B (en) | 2021-06-21 | 2021-06-21 | Character relation mining model training method, character relation mining method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110682820.3A CN113254549B (en) | 2021-06-21 | 2021-06-21 | Character relation mining model training method, character relation mining method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113254549A true CN113254549A (en) | 2021-08-13 |
| CN113254549B CN113254549B (en) | 2021-11-23 |
Family
ID=77188787
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110682820.3A Active CN113254549B (en) | 2021-06-21 | 2021-06-21 | Character relation mining model training method, character relation mining method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113254549B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113806555A (en) * | 2021-09-14 | 2021-12-17 | 国网北京市电力公司 | Operation abnormity identification method, system, device and storage medium for APP |
| CN113946695A (en) * | 2021-12-20 | 2022-01-18 | 山东新希望六和集团有限公司 | Method and device for generating animal pedigree and computer equipment |
| CN113961724A (en) * | 2021-12-22 | 2022-01-21 | 山东新希望六和集团有限公司 | Method and device for updating animal pedigree and computer equipment |
| CN114491080A (en) * | 2022-02-28 | 2022-05-13 | 中国人民解放军国防科技大学 | Unknown entity relationship inference method oriented to character relationship network |
| CN115599927A (en) * | 2022-11-08 | 2023-01-13 | 哈尔滨工业大学（深圳）（哈尔滨工业大学深圳科技创新研究院）（Cn） | Timing sequence knowledge graph completion method and system based on metric learning |
| WO2023155287A1 (en) * | 2022-02-17 | 2023-08-24 | 北京邮电大学 | Complex question-and-answer query method and device based on time sequence knowledge graph |
| CN117114739A (en) * | 2023-09-27 | 2023-11-24 | 数据空间研究院 | Enterprise supply chain information mining method, mining system and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108959418A (en) * | 2018-06-06 | 2018-12-07 | 中国人民解放军国防科技大学 | Character relation extraction method and device, computer device and computer readable storage medium |
| CN110674637A (en) * | 2019-09-06 | 2020-01-10 | 腾讯科技（深圳）有限公司 | Character relation recognition model training method, device, equipment and medium |
| CN111177282A (en) * | 2019-12-30 | 2020-05-19 | 福州大学 | Prediction method for integrating attention mechanism |
| US20210065045A1 (en) * | 2019-08-29 | 2021-03-04 | Accenture Global Solutions Limited | Artificial intelligence (ai) based innovation data processing system |
| CN112765287A (en) * | 2021-02-05 | 2021-05-07 | 中国人民解放军国防科技大学 | Method, device and medium for mining character relation based on knowledge graph embedding |
-
2021
- 2021-06-21 CN CN202110682820.3A patent/CN113254549B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108959418A (en) * | 2018-06-06 | 2018-12-07 | 中国人民解放军国防科技大学 | Character relation extraction method and device, computer device and computer readable storage medium |
| US20210065045A1 (en) * | 2019-08-29 | 2021-03-04 | Accenture Global Solutions Limited | Artificial intelligence (ai) based innovation data processing system |
| CN110674637A (en) * | 2019-09-06 | 2020-01-10 | 腾讯科技（深圳）有限公司 | Character relation recognition model training method, device, equipment and medium |
| CN111177282A (en) * | 2019-12-30 | 2020-05-19 | 福州大学 | Prediction method for integrating attention mechanism |
| CN112765287A (en) * | 2021-02-05 | 2021-05-07 | 中国人民解放军国防科技大学 | Method, device and medium for mining character relation based on knowledge graph embedding |
Non-Patent Citations (1)
| Title |
|---|
| SIXIA MA等: "Learning BiLSTM-based Embeddings for Relation Prediction inTemporal Knowledge Graph", 《JOURNAL OF PHYSICS: CONFERENCE SERIES》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113806555A (en) * | 2021-09-14 | 2021-12-17 | 国网北京市电力公司 | Operation abnormity identification method, system, device and storage medium for APP |
| CN113806555B (en) * | 2021-09-14 | 2023-08-08 | 国网北京市电力公司 | Operation abnormality identification method, system and device for APP and storage medium |
| CN113946695A (en) * | 2021-12-20 | 2022-01-18 | 山东新希望六和集团有限公司 | Method and device for generating animal pedigree and computer equipment |
| CN113961724A (en) * | 2021-12-22 | 2022-01-21 | 山东新希望六和集团有限公司 | Method and device for updating animal pedigree and computer equipment |
| WO2023155287A1 (en) * | 2022-02-17 | 2023-08-24 | 北京邮电大学 | Complex question-and-answer query method and device based on time sequence knowledge graph |
| CN114491080A (en) * | 2022-02-28 | 2022-05-13 | 中国人民解放军国防科技大学 | Unknown entity relationship inference method oriented to character relationship network |
| CN115599927A (en) * | 2022-11-08 | 2023-01-13 | 哈尔滨工业大学（深圳）（哈尔滨工业大学深圳科技创新研究院）（Cn） | Timing sequence knowledge graph completion method and system based on metric learning |
| CN117114739A (en) * | 2023-09-27 | 2023-11-24 | 数据空间研究院 | Enterprise supply chain information mining method, mining system and storage medium |
| CN117114739B (en) * | 2023-09-27 | 2024-05-03 | 数据空间研究院 | Enterprise supply chain information mining method, mining system and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113254549B (en) | 2021-11-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113254549B (en) | Character relation mining model training method, character relation mining method and device | |
| Xiao et al. | Deep salient object detection with dense connections and distraction diagnosis | |
| CN108960409B (en) | Method and device for generating annotation data and computer-readable storage medium | |
| CN111310050B (en) | Recommendation method based on multilayer attention | |
| CN111046275B (en) | User label determining method and device based on artificial intelligence and storage medium | |
| CN111680678B (en) | Target area identification method, device, equipment and readable storage medium | |
| CN110472659B (en) | Data processing method, device, computer readable storage medium and computer equipment | |
| CN110162549B (en) | Fire data analysis method and device, readable storage medium and terminal equipment | |
| CN112328881B (en) | Article recommendation method, device, terminal equipment and storage medium | |
| CN113726545A (en) | Network traffic generation method and device for generating countermeasure network based on knowledge enhancement | |
| CN110851708A (en) | Negative sample extraction method and device, computer equipment and storage medium | |
| Muhammad et al. | A Conceptual Framework for Detecting Learning Style in an Online Education Using Graph Representation Learning | |
| CN107688596B (en) | Burst topic detection method and burst topic detection equipment | |
| CN114417161B (en) | Virtual article time sequence recommendation method, device, medium and equipment based on special-purpose map | |
| Veeravasarapu et al. | Model-driven simulations for computer vision | |
| CN111905375B (en) | Display method, device and related equipment | |
| CN115033700A (en) | Cross-domain emotion analysis method, device and equipment based on mutual learning network | |
| CN112507214B (en) | User name-based data processing method, device, equipment and medium | |
| CN116764236A (en) | Game prop recommending method, game prop recommending device, computer equipment and storage medium | |
| CN110232393B (en) | Data processing method and device, storage medium and electronic device | |
| CN113392309A (en) | Data processing method, device, equipment and storage medium | |
| JP6954070B2 (en) | Discrimination program, discrimination method and discrimination device | |
| CN113282740B (en) | Content recommendation method and device, electronic equipment and storage medium | |
| CN114298118B (en) | Data processing method based on deep learning, related equipment and storage medium | |
| CN115658878B (en) | Intelligent fishery knowledge recommendation method, system and medium based on knowledge graph |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |